Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/85612
DC FieldValueLanguage
dc.contributorDepartment of Computing-
dc.creatorLin, Wan-kei Wilfred-
dc.identifier.urihttps://theses.lib.polyu.edu.hk/handle/200/1487-
dc.language.isoEnglish-
dc.titleSoft computing based dynamic buffer tuning for better response timeliness and fault tolerance for Internet channels-
dc.typeThesis-
dcterms.abstractThe area of this PhD research is directed towards performance enhancement and fault-tolerance at client/server(C/S) interaction over a logical Internet channel. The aim is to effectively eliminate the user-level buffer overflow so that retransmissions can be reduced to shorten the service roundtrip time (RTT) in the interaction. Since a server may serve different clients simultaneously, the relationship is actually one-server-to-many-clients, alternatively known as the asymmetric rendezvous. The different streams of service requests from clients merge at the server's queue and this easily inundates the queue buffer to overflow at peak times. In fact, an asymmetric rendezvous involves two levels: the system/router level that includes all activities inside the TCP channel, and the user level that involves the client and the server. If the collective error probability for a client/server interaction path is Ppath , then the average number of trials (ANT) to send a message successfully from one end of the C/S path to another is ANT= k->8 E j=1 jP j-1 path (1-Ppath) ~~ 1/(1-Ppath). Since Ppath also encapsulates the user-level buffer overflow error, eliminating the latter definitely yields a smaller ANT and shorter end-to-end service roundtrip time (RTT). My previous MPhil research concluded that dynamic buffer size tuning can indeed eliminate the chance of user-level buffer overflow. This was clearly demonstrated by the experimental results with the dynamic buffer controllers proposed. These original controllers developed in the MPhil thesis are: 1) PIDC ("proportional (P) + integral(I) + derivative(D)" Controller): It is algorithmic and always eliminates user-level buffer overflow but has two shortcomings: a) it locks unused memory, and b) it does not have a safety margin and therefore the queue length can get dangerously close to the buffer length, threatening possible overflow. 2) GAC (Genetic Algorithm Controller): It is the "PIDC + genetic algorithm (GA) + {0,A}2 objective function" combination. The GA moderates the PIDC process so that the outcome is always within the +-A safety margins about the steady-state reference symbolically represented by "0" in {0,A}2 . The GA eliminates the PIDC shortcomings but also produces occasional buffer overflow because it does not guarantee the global-optimal solution of the solution hyper-plane. 3) FLC (Fuzzy Logic Controller): It is the combination: "PIDC + fuzzy logic + {0,A}2 objective function" combination, which was proposed to preserve the GAC merits and eliminate the occasional buffer overflow. The fuzzy logic moderates the PIDC control process similar to the GA. 4) NNC (Neural Network Controller): It works with the {0,A}2 objective function but does not include PIDC. Its proposal was inspired by the successful experience of using neural networks in AQM (active queue management) algorithms, which prevent network congestion at the system/router level. AQM methods differ from the dynamic buffer size tuners by using a fixed-size buffer. When experiments were conducted to verify the above four dynamic buffer tuners, it was observed that their performance was affected by the traffic patterns. The conclusion is that measures must be taken to neutralize the ill effects by traffic on tuner stability and accuracy. My MPhil thesis left several unaddressed issues that form the backbone of this PhD research. The issues include: 1) In the aspect of traffic ill effects: a) Is it possible to calibrate the ill effects offline so that the tuners can use these calibrations to ward off traffic changes by fine-tuning its dynamic buffer tuning process adaptively? b) If so, then how can the current Internet traffic pattern be deciphered on the fly (on-line) so that the off-line calibrations can be applied selectively? 2) For FLC: a) Is it possible to have an optimal design? b) Is it possible to make the tuner self-reconfigurable (especially with respect to traffic pattern changes)? 3) For NNC: a) Is it possible to prune the NNC configuration on the fly so that its control cycle time can be consistently and adaptively reduced? b) Is there a correlation between control accuracy and the number of hidden neurons in the NNC back-propagation architecture? (The procedure to provide the answer is called sensitivity analysis.) The motivation of my PhD research is to provide answers to the above unaddressed issues. As a result the following solutions are proposed: 1) For real time traffic analysis: Two traffic filters have been proposed: real-time modified QQ-plot (or simply RT-QQ) and self-similarity ( S2 ) filter. These filters identify the Internet traffic patterns on the fly. The RT-QQ recognizes heavy-tailed distributions and the S2 filter identifies self-similarity. 2) For FLC: a) an optimal design range is found for FLC design, and b) a way is found to make the FLC adaptive/reconfigurable by squeezing the "don't care" state range threshold in a dynamic manner. 3) For NNC: a) the HBP (Hessian Based Pruning) approach was proposed for pruning or optimizing the NNC configuration on the fly and as a result its average execution time (i.e. control cycle time) is reduced, and b) sensitivity analysis was conducted and the results confirm that more hidden neurons do not necessarily mean better NNC performance. The solutions proposed in my PhD research have contributed to 19 publications so far (5 journals and 14 conferences). All the stated PhD research objectives have been achieved. The research has also uncovered many relevant problems, which should be resolved in the future work: a) investigation of the issue of how to choose the limits for Gaussian tests effectively, b) deepening of the investigation into why "heavy-tailedness" is not a necessary condition of self-similarity, and c) investigation into how the dynamic buffer size controllers, especially the FLC, can best support pervasive computing based e-applications such a telemedicine.-
dcterms.accessRightsopen access-
dcterms.educationLevelPh.D.-
dcterms.extent260 leaves : ill. ; 30 cm.-
dcterms.issued2005-
dcterms.LCSHHong Kong Polytechnic University -- Dissertations.-
dcterms.LCSHInternet.-
dcterms.LCSHInternetworking (Telecommunication)-
dcterms.LCSHComputer system failures.-
dcterms.LCSHFault-tolerant computing.-
Appears in Collections:Thesis
Show simple item record

Page views

49
Last Week
0
Last month
Citations as of Mar 24, 2024

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.